AI Observability
AI Observability overview
Track AI requests, traces, providers, models, tokens, latency, errors, alerts, and estimated cost.
AI Observability is for teams that need to understand how their AI systems behave in production.
What it tracks
- Provider, model, environment, service, endpoint, and status.
- Input tokens, output tokens, total tokens, latency, and estimated cost.
- Trace groups, request detail, spans, metadata, and errors.
- Model performance, usage trends, prompt patterns, routing options, alerts, and evaluations.
Main tabs
- Overview for the high-level picture.
- Traces and Request Detail for debugging one flow.
- Cost and Models for spend and performance.
- Errors and Alerts for reliability problems.
- Playground and Evaluations for prompt and model testing.
- Scores, Sessions, and Users for quality operations.
- Recommendations and Prompts for optimization.
AI providers and gateways
- OpenAI: usage ingestion, provider key status, logs, per-key usage, playground execution, traces, model analytics, cost analytics, errors, and pricing lookup.
- Anthropic: usage ingestion, provider key status, playground execution, traces, model analytics, cost analytics, and error tracking.
- Gemini: usage ingestion, provider key status, playground execution, traces, model analytics, cost analytics, and error tracking.
- Amazon Bedrock: AWS-first metrics, Bedrock console panel, Bedrock sync, traces, cost, model usage, errors, and provider metadata.
- Custom endpoints: trace and event ingestion for self-hosted or gateway-based models when telemetry is instrumented.
- Direct SDK Ingestion: lightweight client SDKs that record provider, model, request, usage, latency, cost, and error metadata directly from the source.
Need more help?
Keep moving with the right next guide
If your team still runs into setup issues, empty dashboards, billing delays, callback failures, or alerting problems, continue with troubleshooting before re-running the entire onboarding flow.